This is the second post in the series I am calling What I Measured. The rule is that nothing goes in unless I counted it, and the count gets published whichever way it falls. Here it falls badly ten times.
One caveat first, because leaving it out would be the same dishonesty I am writing about. These ten are selected precisely because measurement contradicted them. I held plenty of other beliefs that measurement confirmed, and those are boring. So the score in the title is not evidence that everything I think is wrong. It is a list of the ten places where I was confident and the number disagreed.
What did I actually measure, and how?
Everything below is either a count taken on my own two sites, or a direct statement from Google’s own documentation that contradicts something I had been told by the wider SEO industry. No third-party estimates, no tool scores, no vendor dashboards. Where a number appears, it is a floor from a source I can re-run, not an estimate. Where a belief was killed by documentation rather than by a count, I say so and link the document rather than paraphrasing it.
Which beliefs about getting found were wrong?
These four are about the basic mechanics of Google finding and ranking a page. All four are things I would have said out loud with confidence a year ago.
1. If the writing is good enough, links show up on their own. I audited my own inbound links across two sites and found exactly one. Not one good one. One. It came from a directory I never submitted to, it listed the wrong city, and it pointed at a URL that is not my canonical, which I wrote up in the backlink audit that turned up a single wrong link. Quality does not generate links. Asking generates links.
2. Pages that are not indexed have been judged and rejected. Wrong, and the distinction matters enormously. At one point 15 posts across my two sites were not in Google’s index, and I assumed a quality problem. Search Console reported zero URLs under “crawled, currently not indexed”. Zero. That is not rejection. That is non-discovery. Robots, canonicals, sitemap and server-rendered links all checked out. Google had simply never come. Which loops straight back to finding one, because Google finds pages mostly through links, and I had a total of one.
3. Topical authority is a score I can raise by publishing more. Google ships no topical authority score, and a Search Advocate has publicly called the idea relevance under a different name. The leaked Content Warehouse documentation does list site-level focus and radius fields, which is why the industry belief is not pure invention. Both things being true at once is the actual answer, and I worked through it in the post on what topical authority actually is. What I had been doing, publishing widely and calling it authority building, was making the measured thing worse.
4. E-E-A-T is a ranking factor I should optimise for. Google’s own words on this are blunt. Their SEO Starter Guide list of things not to focus on answers the question “is E-E-A-T a ranking factor” with “no, it’s not”. The same list kills three more things I had been treating as real: there is no magic word count, there is no duplicate-content penalty, and the meta keywords tag is unused. I had shipped audit findings built on at least two of those.
Which beliefs about AI search were wrong?
Two of the ten are about getting cited by AI answers, which is where the largest gap sits between what vendors sell and what the platform documents.
5. An llms.txt file gets me into Google’s AI answers. It does not. Google’s guide to optimising for AI features in Search says plainly that Google Search does not use llms.txt, that chunking content for AI is unnecessary, and that rewriting content “for AI” is pointless because the systems already understand meaning. I had all three of those on a services page as though they were levers. They may still matter for other engines, and I now label them that way instead of selling them as Google work.
6. A higher Lighthouse score is a better site. Two measurements broke this. First, a local build reported mobile scores that the live deployed URL flatly contradicted, so half my performance work was chasing a number that only existed on my laptop. Second, and worse, when the score went up the amount of content a no-JavaScript crawler could see went down, because the wins came from moving things client-side. That is the same shape as the day I found the AI step that ran fine and still cost me time: the metric improved, the thing the metric was a proxy for got worse.
Which beliefs about my own tooling were wrong?
The last four are about the machinery I use to do the work, and they are on this list because unreliable tooling produces unreliable SEO conclusions.
7. Installing the backlink tool means I can see my backlinks. The tool referenced five scripts across fourteen lines of its own instructions. Zero of them existed on disk. I also spent an afternoon on a public web-graph dataset before establishing that the file I was downloading contains a centrality rank per domain and no referring-link edges at all, so it can never answer “who links to me”. This is the pattern from the tool I bought and never opened again, except worse, because I nearly reported numbers from it.
8. A green audit score means the page is fine. Aggregate scores hide the cases that matter. I only ever found real regressions by reading the individual outputs rather than the summary number, which is the whole argument in why I stopped trusting eval scores and started reading transcripts. An SEO score behaves identically. It passes on average while the one page that carries your revenue quietly fails.
9. My publishing automation is working because it exits clean. It ran for two weeks, returned success every time, and produced nothing usable, which I documented in the automation that failed for two weeks without throwing an error. Exit code zero means the process finished. It says nothing about whether the outcome happened. Every scheduled SEO job I run now checks the outcome, not the exit code.
10. Installing a capability is the same as having it. I counted 260 skills installed on my machine against five months of session logs. 150 had never run once, and only 36 had run in more than one session. I published the full ranking, losers included, in the count of which installed skills have actually fired, the sibling post to this one. The install list was never the capability list.
What did all ten have in common?
Every one of these ten beliefs survived for months because nothing threw an error. There was no alert for zero backlinks, no warning that the posts had never been discovered, no failure when a tool referenced scripts that did not exist. Wrong SEO beliefs do not announce themselves. They sit there being plausible until somebody runs a count against them, and the count is almost always cheaper than the belief was.
The practical version of this, if you take nothing else: pick the three things you are most confident about in your own marketing and find the smallest number that could contradict each one. Not a dashboard. A number you can compute yourself and re-run next month. For me those three were inbound links, indexed pages, and whether the scheduled job produced anything. Two of the three were at zero and I had no idea.
That is most of what I actually do when a founder hires me for AI and search work. Not adding tactics. Finding the two or three confident assumptions holding up the whole plan and putting a number next to each one, because the tactics built on a wrong assumption are the expensive part.
Ten beliefs. Ten counts. Ten corrections. The next ten will not be any more flattering, and I will publish those too.